History and Future of Autonomous Systems

Click any point on the line. Each one is a moment where the loop between sensing and acting got a little more automatic, and a little less dependent on a person standing in the middle of it.

Autonomy didn't start with computers. James Watt's centrifugal governor, patented in 1788, kept a steam engine at a constant speed with no logic circuit at all, spinning weights physically pulled a throttle closed as engine speed rose and let it open as speed fell. It's a feedback loop built entirely out of metal.

The theory that gave this loop a name and a general mathematics came a century and a half later. Norbert Wiener's 1948 book Cybernetics treated animals and machines as the same kind of system: something that senses an outcome, compares it against a goal, and corrects its next action based on the difference (Wiener, 1948). That closed loop, sense, compare, correct, act, is still the skeleton every framework on this timeline builds on, including the perceive-plan-act-observe loop this series covers in its treatment of modern AI agents.

George Devol patented a programmable robotic arm in 1954, and in 1961 the first Unimate went to work on a General Motors assembly line in Trenton, New Jersey, lifting die-cast metal parts too hot for a person to handle safely. It followed a fixed, memorized sequence, not a sensed one, closer to a player piano than a feedback loop, but it proved a machine could be reprogrammed for a new job rather than rebuilt for one.

Shakey, built at Stanford Research Institute between 1966 and 1972, was the first robot to close the whole loop with reasoning in the middle: it used a camera and sensors to build a model of a room, planned a sequence of actions to reach a goal, and moved through the room executing that plan, adjusting when the plan didn't match what it sensed. It ran on a mainframe the size of a room to do what a modern AI agent does with a perceive-plan-act-observe cycle, the same shape, decades earlier and orders of magnitude slower.

The honest version of this history isn't a smooth line upward. Optimism about machine reasoning ran far ahead of what early systems could actually deliver, twice, and funding collapsed both times when the gap became impossible to ignore, roughly the mid-1970s and again in the late 1980s into the 1990s. Expert systems, the commercial hope of the second boom, turned out to be expensive to build and brittle to maintain once they left the narrow domain they were tuned for.

DARPA's 2004 Grand Challenge offered a million dollars to the first team whose vehicle could autonomously drive 142 miles across the Mojave Desert. Fifteen vehicles started. None finished. The best of them, CMU's entry, made it 7.4 miles before getting stuck against a rock. It was a public, embarrassing failure, and DARPA ran it again eighteen months later anyway.

In October 2005, five vehicles finished the full course. Stanford's "Stanley," a modified Volkswagen Touareg led by Sebastian Thrun, won in just under seven hours, using sensor fusion and probabilistic reasoning rather than a fixed script (DARPA, 2014). Two years later, the 2007 Urban Challenge moved the test into a mock city with traffic, intersections, and other moving vehicles, and multiple teams finished that too.

Three events, seven years apart, reset what people thought autonomous systems could learn to do on their own rather than be explicitly programmed to do. In 2012, a neural network called AlexNet beat every prior approach at the ImageNet image recognition competition by a margin nobody expected, proving that deep, GPU-trained networks could outperform hand-engineered vision systems. In March 2016, DeepMind's AlphaGo defeated Lee Sedol, one of the strongest Go players alive, in a game experts had assumed was at least a decade beyond machine reach because of how much it rewarded intuition over brute-force search.

The architectural shift that made the current era possible came in 2017. Vaswani and colleagues' paper "Attention Is All You Need" introduced the transformer, a way of processing sequences that scaled dramatically better than what came before, and it is now the foundation underneath essentially every large language model in use (Vaswani et al., 2017).

Once transformers made large language models fluent, the next question was whether they could do more than answer in place: could a model itself run the sense-compare-correct-act loop from Sheet 2, using its own reasoning to decide what to do next? Yao and colleagues' 2023 ReAct paper showed a language model interleaving reasoning and action, deciding what to check, checking it, and updating its plan based on what came back, without a person scripting each step (Yao et al., 2023). A year later, Anthropic's Model Context Protocol gave agents a standard way to actually reach the tools they needed to act on, the wiring behind the loop, not just the reasoning inside it (Anthropic, 2024).

Between them, these two pieces are the direct descendants of Shakey's camera-plan-move cycle from 1970, running now on language instead of a room-sized mainframe, and fast enough to matter.

The most concrete, evidence-based claim currently on record about where this is headed comes from METR, a research organization that measures how long a task an AI agent can complete on its own with 50% reliability, using the length of time a skilled human would need as the yardstick. Tracking this across six years of frontier models, they found the number has been doubling roughly every seven months since 2019, reaching just under an hour for the best model the paper measured in early 2025 (Kwa et al., 2025). Continued tracking on the same measure has reported substantially longer horizons since, reportedly past sixteen hours by 2026, though those later figures come from ongoing measurement rather than that paper.

What this trend doesn't say matters as much as what it does. It's a measurement of one specific thing, task length on a particular benchmark suite, extrapolated forward, and METR's own paper spends real space on the limits of that extrapolation: whether the trend holds outside the tasks it was measured on, and whether recent years are running faster than the six-year average. Every framework covered across this series, situation awareness, calibrated trust, the human-in-the-loop checkpoint, gets more consequential, not less, as the length of task a system can run unsupervised keeps climbing. A longer leash is exactly the condition under which losing track of what's on the other end of it matters most.